Propose before you apply
Automated cleanup that writes directly to the CRM will eventually merge two companies that were genuinely different or overwrite a field someone maintained by hand, and the damage is discovered weeks later without a way to reverse it. The safer shape proposes changes to the record owner with the evidence attached and applies only what is approved, moving to automatic application per rule once that rule has demonstrated it is boring. Deletion stays manual permanently. A change log covering every automated edit is what makes the whole thing auditable rather than mysterious.
Matching is probabilistic, so plan for the middle
Duplicate detection produces three groups: obvious matches, obvious non-matches, and a middle band that no threshold resolves cleanly. Tuning to eliminate that middle band only moves errors into the automatic groups where nobody sees them. We size the review queue deliberately, keep the middle band visible, and make working through it a short scheduled task rather than a project. Source-of-truth rules decide which system wins on conflict, and enrichment data carries its collection date, since a job title from three years ago is not a fact — it is a guess with a timestamp.
Pilot CRM cleanup with reversible proposals
For a company-record review, compare stable identifiers first, then treat normalized name and domain as evidence rather than proof of a match. Produce a proposed duplicate pair with the fields, source and match reasons visible; keep uncertain pairs in a review queue and never merge from a fuzzy score alone. Apply only owner-approved field changes, retain the previous values and provenance, and keep deletion manual. Use a sample labeled by CRM owners to measure confirmed-match precision, reviewer overrides, queue size and field correction rate; set acceptance before enabling any automatic rule, and verify the rollback path against the change log.
How to measure the pilot
A pilot is accepted against the current process, not an industry promise. Agree the sample, baseline and owner before launch, then review the same measures after the workflow has handled real cases.